Three years on since ChatGPT unleashed itself onto the global stage, and the world is still learning how to keep up with the breakneck speed of AI innovation. Every aspect of AI is fiercely debated, yet, at the same time, every stakeholder of this breathtaking technology agrees on one thing—it is not just change but rapid change that is the only constant that matters in AI.
Take 2025 for example. ‘Model inference’ and ‘AI bubbles’ were the two main topics that furiously dominated the investor’s playbook in AI after two years of rapid change, intense focus, and ebullience surrounding ‘model training’. But, while AI bubbles seem to have sort of taken a back seat (yet again), the debate around model inference continues to rage.
Suddenly, shares of old boring semi businesses like memory and storage manufacturers surged over the past few months ( more precisely in the year 2025).
In this post, the Pragmatic Engineer newsletter re-evaluate their views on the market’s new shopping basket favorite, Micron $MU , while also establishing their views on the data center storage market, which involves companies like Seagate $STX , Sandisk $SNDK , Western Digital $WDC , and Pure Storage $PSTG .
Additionally, the Pragmatic Engineer newsletter believe this hype and exuberance surrounding the astronomical surge in memory prices will uniquely benefit Astera Labs $ALAB and Marvell Technologies $MRVL .
Brief Overview Of AI Data Center’s Memory & Storage Ecosystem
Until 2025, markets were honeymooning with AI and rewarding companies for model performance, which often put focus on the trillions of tokens that these models were trained on. As a result, GPUs and other forms of compute were being thrown at these models to power model training workloads.
Then, through 2025, markets gravitated towards assessing the degree of monetization of these models, which shifted the focus from model training towards model inference, silently beginning the shift in some of the capex budget dollars towards memory and storage in DCs (data centers).
Additionally, over the past year, context windows for AI models, especially for frontier models like most of Google’s Gemini models, exploded, as seen below.
(Context windows are like the working memory of an AI model, where a user’s previous chats, documents, images, videos, etc., are stored, and the model recalls memory from its context windows, enabling it to remember long, fluid conversations users have had with it.)
Also until last year (2025), a model’s context window was typically held in HBM, or high bandwidth memory—the 4, 8, or 12-high stacks and the recently launched 16-high stacks of packaged DRAM that became the revolutionary face of memory makers like Micron.
If HBMs were not enough during a context window recall, a model would perform the recall from additional DRAM that was stacked alongside the CPU. But the belligerent expansion of model context windows and the rapidly scaling ability of models to go multi-modal through last year forced Data center architects to add additional storage options for AI models, which took the form of SSDs, or solid-state drives, and HDDs, or hard disk drives. This pushed up the demand for SSDs and HDDs.
So while 2025 did start with Micron being a strong contender due to the severe demand for HBM, the growing need for server storage also propelled shares of Sandisk, Seagate, and Sandisk, to name a few, alongside Micron, creating a new breed of semiconductor winners in 2025.
The Views of the Pragmatic Engineer newsletter On Micron’s market storage ecosystem .
The Pragmatic Engineer newsletter anticipated a big move in Micron in 2025 and took a starter position in the American HBM maker’s shares as part of their legacy portfolio at the start of last year ( 2024). While they closed out their position with a cumulative gain of +130%, with their last tranche sold in December 2025. They have used the past couple of weeks to re-establish their views on the company, given its fiercely relentless momentum. Two factors helped them change their views.
The first factor was already known to them because they did believe in Micron’s early commitment to capture 25% of the HBM market when its management had staked claims on the fledgling HBM market in 2024. 15 months later, Micron has beautifully tracked towards its own goals to perfection based on recent market data and should close CY25 with a market share just above 25%.
This was only possible because Micron moved with agility to market their HBM product, packaging and validating their HBM designs with major GPU/XPU companies like Nvidia, etc. Per Micron’s comments from last month’s ER call, the company should close out CY25 with bit shipment growth in the “low 20% range.”
This brings them to the second factor. On the same ER call last month (December 2025), Micron’s time-tested management revealed that the company would be targeting ~20% overall bit shipment growth in CY26 across all its memory products, marginally below their CY25 forecast.
In conclusion we can see that the storage ecosystem is a critical part of the AI innovation. That is why engineers are now thinking about « orbital data centers ».
